3D Reconstruction Based On Mobile RGB-D Camera

The three dimensional(3D)reconstruction is a hot topic in computer vision and graphics.For decades,many excellent methods of reconstruction have emerged.According to the different methods of data collection,these methods can be divided into the following categories:the 3D reconstruction based on visual image,the 3D reconstruction based on laser scanner and the 3D reconstruction based on RGB-D camera.The 3D reconstruction technology based on laser scanner needs to set the target in the scene in advance,and the laser scanner is expensive and difficult to popularize.The 3D reconstruction technology based on visual image can recover the 3D model of the scene by the two-dimensional image which captured by a common digital camera,the rich visual information can estimate the pose of the camera very well.However,this method is not effective in the case of texture information insufficient.With the development of 3D imaging technology and the advent of consumer RGB-D depth cameras,the 3D reconstruction technology based on RGB-D has been rapidly developed.Benefited from the RGB-D depth camera can obtain the texture and geometric information of the scene simultaneously with great portability and low cost.In this paper,the 3D reconstruction method based on RGB-D depth camera is studied.The typical 3D reconstruction methods based on RGB-D depth camera include KinectFusion,Kintinuous,and ElasticReconstruction.The above-mentioned algorithms can achieve satisfactory reconstruction results in an ideal environment.Kintinuous algorithm solves the shortcomings of KinectFusion algorithm through scalable TSDF mesh model,but the point cloud slicing stream extracted from the model cannot participate in subsequent pose optimization.The ElasticReconstruction algorithm only optimizes the pose of the segmented local scene model,and the reconstruction quality is influenced by the pose estimation of the inner image in local scene model.The main research work and innovation are summarized as follows:The pose estimation based on texture features and geometric information is studied.Firstly,the statistical histogram of geometric information such as normal vectors and the SIFT texture feature are combined in the process of feature extraction,which improves the discrimination ability of feature descriptors under various conditions such as illumination,rotation and scale.Secondly,according to the initial registration of feature points based on feature descriptors,the relative transformation matrix between matching points and the matching results with high accuracy are calculated by using the RANSAC algorithm and the multiple error constraints of matching points.Finally,the validity and robustness of the proposed method are verified by comparing with SIFT feature and EPnP algorithm in three open datasets.The pose optimization method based on graph model is studied.The pose optimization algorithm in ORB-SLAM algorithm is improved,the key frame generation rules are modified to solve the solution of tracking failure,and the pose optimization module of the scene frame is added.In order to improve the robustness of the tracking process,the bundle adjustment is used to optimize the estimated pose of the scene frame.